AMD Wants to Know If You'd Like Ryzen AI Support on Linux | Hacker News Reader

AMD Wants to Know If You'd Like Ryzen AI Support on Linux

phoronix.com

56 pointsclouddrover21 comments

The important part: let them know on this GitHub issue by adding a comment that says +1:

https://github.com/amd/RyzenAI-SW/issues/2

And yes, it’s weird that they have to ask. I thought a significant amount of AI work was done in Linux.

The bulk of local Stable Diffusion and Llama usage seems to be on Windows.

This is not my scientific assessment, but a very strong feel from looking at Github issues, Discord chat and AI Horde hosts.

Even devs for some of the big frameworks seem to be primarily Windows based.

It's not like actual AI dev will take place on ryzenai hardware. At best it would be for accelerating current workloads. It's very very... lackluster in terms of performance even for basic inference tasks

Better yet, +1 the comment by the AMD employee who specifically asked for it: https://github.com/amd/RyzenAI-SW/issues/2#issuecomment-1773...

I mean, it barely even exists or seems to be acknowledged in Windows either. Somehow, they made it into a feature that OEMs can decide to disable , even with compatible CPUs. You could buy a CPU marketed as having ryzenai without being able to actually use it. This issue (and the complete lack of info, save from the input of a very helpful employee) to me shows the state of ryzenai:

https://github.com/amd/RyzenAI-SW/issues/5#issuecomment-1726...

If AMD does not improve the quality of their drivers and ML stack, they can kiss the ML market goodbye. I have tended to buy the best hardware per dollar for the last 20 years. Sometimes I have Intel CPUs, sometimes AMD, sometimes both, but I stopped buying AMD GPUs in 2007. I completely switched to Linux from FreeBSD in 2003, and the only viable option for me has always been NVIDIA due to the availability and quality of their drivers. I don't mind AMD losing the ML race, though it is very sad that NVIDIA does not have strong competitors.

I'd just be happy if nested virtualization worked in windows for vmware workstation...

So my linux VM can run kvm.

Just release sufficient documentation, and those motivated enough will write the necessary drivers.

I don't want any AI on any chipset. At the very least, there should be an option to disable hardware and software based AI.

Why does it matter if its not being used?

I think the die area for AMD's ASIC is pretty small.

I'd rather not pay extra for silicon I could have someone design something more useful for.

Eh, honestly, the "AI" selling point alone probably reduces the cost of the die though amortization/more sales more than leaving it out would.

I have some quips about Phoenix, but the tiny AI block with poor software support is very low on the list.

true, I've heard about "dark silicon" and "binning", although some of that is due to saving on power rather than selling different product segments https://semiengineering.com/knowledge_centers/low-power/tech...

And what will this AI support mean ? Better branch prediction ?

Do they think it's optional or something?

I'd expect Linux to be the main platform, is AMD trying to sabotage itself?

Secret deal with Microsoft perhaps?

I'm sure they just want to give Nvidia a fighting chance in the AI space, since Nvidia is so far behind.

The AI engines on these chips are derived from the Xilinx AIE / AIE-ML cores. The Xilinx aiecompiler only runs on Linux - which is special because, aside from aiecompiler, basically all of the Xilinx tools work on both Windows and Linux.

In contrast the Ryzen AI stuff is more consumer focused. And 99% of consumers are using Windows. Having Linux support would be nice, but it totally makes sense that they would implement it on Windows first and prioritize that. Linux probably isn't worth it aside from for developer mindshare. And I say this as someone who's been using desktop Linux for over a decade.

If they have to meekly ask if anyone wants linux support... Usage not going to proliferate anyway.

Getting support from popular projects for Metal, ROCm, Vulkan, Intel's frameworks and Nvidia TensorRT is already like pulling teeth, even with Intel, Apple and Nvidia paying engineers to integrate the already functional frameworks into projects.

The question should be if AMD thinks they sell enough of these Xilinx chips to support.

Do they not see the piles of money Nvidia is making? Support for all of the major platforms should be pretty obvious from their position

AMD messes up software... Again.

Reply on news.ycombinator.com